Csound › Csound - Dev

Re: [Csnd] scons MSVC report

Classic

List

Threaded

3 messages Options

Istvan Varga

Re: [Csnd] scons MSVC report

Victor Lazzarini wrote:
> What's the assembly code then for truncation
>
> i = (int) f;
>
> for double & single precision?
>

Well, compiling this C code with gcc 4.0:

/* -------- start test.c -------- */

#define _ISOC99_SOURCE
#include <math.h>

int f2i_cast(float x)
{
return (int) x;
}

int d2i_cast(double x)
{
return (int) x;
}

int f2i_lrint(float x)
{
return (int) lrintf(x);
}

int d2i_lrint(double x)
{
return (int) lrint(x);
}

/* -------- end test.c -------- */

For a generic x86 CPU (irrelevant lines filtered out):

Compiler flags: -Wall -O2 -fomit-frame-pointer -S -masm=intel

f2i_cast:
sub %esp, 8
fnstcw WORD PTR [%esp+6]
fld DWORD PTR [%esp+12]
movzx %eax, WORD PTR [%esp+6]
or %ax, 3072
mov WORD PTR [%esp+4], %ax
fldcw WORD PTR [%esp+4]
fistp DWORD PTR [%esp]
fldcw WORD PTR [%esp+6]
mov %eax, DWORD PTR [%esp]
add %esp, 8
ret

d2i_cast:
sub %esp, 8
fnstcw WORD PTR [%esp+6]
fld QWORD PTR [%esp+12]
movzx %eax, WORD PTR [%esp+6]
or %ax, 3072
mov WORD PTR [%esp+4], %ax
fldcw WORD PTR [%esp+4]
fistp DWORD PTR [%esp]
fldcw WORD PTR [%esp+6]
mov %eax, DWORD PTR [%esp]
add %esp, 8
ret

f2i_lrint:
sub %esp, 16
fld DWORD PTR [%esp+20]
#APP
fistpl DWORD PTR [%esp+12]
#NO_APP
mov %eax, DWORD PTR [%esp+12]
add %esp, 16
ret

d2i_lrint:
sub %esp, 16
fld QWORD PTR [%esp+20]
#APP
fistpl DWORD PTR [%esp+12]
#NO_APP
mov %eax, DWORD PTR [%esp+12]
add %esp, 16
ret

Now for Pentium III, using SSE 1 instructions:
Compiler flags: -Wall -O3 -march=pentium3 -fomit-frame-pointer -ffast-math -S -masm=intel

f2i_cast:
cvttss2si %eax, DWORD PTR [%esp+4]
ret

(the other functions do not change significantly)

Now try Pentium 4, with SSE 2:
Compiler flags: -Wall -O3 -march=pentium4 -fomit-frame-pointer -ffast-math -S -masm=intel

f2i_cast:
cvttss2si %eax, DWORD PTR [%esp+4]
ret

d2i_cast:
cvttsd2si %eax, QWORD PTR [%esp+4]
ret

(again, the lrint based functions remain effectively the same)

As you can see, enabling the use of SSE (with -march) can significantly
improve float to integer casts.

-------------------------------------------------------
SF.Net email is sponsored by: GoToMeeting - the easiest way to collaborate
online with coworkers and clients while avoiding the high cost of travel and
communications. There is no equipment to buy and you can meet as often as
you want. Try it free.http://ads.osdn.com/?ad_id=7402&alloc_id=16135&op=click
_______________________________________________
Csound-devel mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/csound-devel

Remove Ads

Istvan Varga

Re: [Csnd] scons MSVC report

Victor Lazzarini wrote:

> That code is for rounding isn't it? What about a version for truncation,
> which is used a lot.

For truncating floats (not doubles) to 32 bit integers, you can try this
(perhaps it may be improved further, but first tests show that it is faster
than casting - at least with my CPU - even if not inlined):

/* uncomment these for MSVC */
/*
typedef unsigned char uint8_t;
typedef int int32_t;
typedef unsigned int uint32_t;
#define inline __inline
*/

static inline int32_t float2int32(float x)
{
union {
float f;
int32_t i;
} u;
uint32_t tmp;
uint8_t tmp2;
u.f = x;
tmp2 = (uint8_t) 158 - (uint8_t) (((int) u.i & 0x7F800000) >> 23);
if (tmp2 & (uint8_t) 0xE0)
return (int32_t) 0;
tmp = (uint32_t) u.i | (uint32_t) 0xFF800000UL;
tmp = (tmp << 8) >> tmp2;
return (u.i < (int32_t) 0 ? -((int32_t) tmp) : (int32_t) tmp);
}

Some benchmark results with the following test code:

volatile int32_t foo;

int main(int argc, char **argv)
{
float tmp = -12345678;
while (tmp < 12345678) {
foo = float2int32(tmp);
tmp += 0.7071;
}
return 0;
}

cast, no SSE: 0.996
case, SSE: 0.164
float2int32 (inline): 0.356
float2int32 (not inline): 0.550

Here is the assembly listing of float2int32 generated with GCC for a
Pentium Pro or newer CPU (as uses cmov):

float2int32:
push %ebx
mov %al, -98
mov %ebx, DWORD PTR [%esp+8]
mov %edx, %ebx
shr %edx, 23
sub %al, %dl
movzx %edx, %al
xor %eax, %eax
test %dl, -32
jne .L4
mov %eax, %ebx
mov %cl, %dl
or %eax, -8388608
sal %eax, 8
shr %eax, %cl
mov %edx, %eax
neg %edx
inc %ebx
cmovle %eax, %edx
.L4:
pop %ebx
ret

-------------------------------------------------------
This SF.Net email is sponsored by Yahoo.
Introducing Yahoo! Search Developer Network - Create apps using Yahoo!
Search APIs Find out how you can build Yahoo! directly into your own
Applications - visit http://developer.yahoo.net/?fr=offad-ysdn-ostg-q22005
_______________________________________________
Csound-devel mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/csound-devel

Istvan Varga

Re: [Csnd] scons MSVC report

Istvan Varga wrote:

> cast, no SSE: 0.996
> case, SSE: 0.164
> float2int32 (inline): 0.356
> float2int32 (not inline): 0.550

Another test with using lrintf, but with hacks for conversion with truncating:

static inline int32_t float2int32(float x)
{
return (int32_t) lrintf(x + (x < 0.0f ? 0.5f : -0.5f));
}

inline: 0.277
if not inlined: 0.502
(compare with above)

assembly (probably needs to be converted to MSVC format if you want to use it)
for Pentium Pro or newer:

.LC1:
.long 1056964608
.LC2:
.long -1090519040

float2int32:
sub %esp, 16
fld DWORD PTR [%esp+20]
fldz
fcomip %st, %st(1)
fld DWORD PTR .LC2
fld DWORD PTR .LC1
fcmovbe %st, %st(1)
fstp %st(1)
faddp %st(1), %st
#APP
fistpl DWORD PTR [%esp+12]
#NO_APP
mov %eax, DWORD PTR [%esp+12]
add %esp, 16
ret

-------------------------------------------------------
This SF.Net email is sponsored by Yahoo.
Introducing Yahoo! Search Developer Network - Create apps using Yahoo!
Search APIs Find out how you can build Yahoo! directly into your own
Applications - visit http://developer.yahoo.net/?fr=offad-ysdn-ostg-q22005
_______________________________________________
Csound-devel mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/csound-devel